Programmable linear-optical interferometer (no entangling primitive) g_lointer

A passive network maps input modes onto outputs, a†ᵢ → Σⱼ Uⱼᵢ a†ⱼ, with U an N×N unitary. Reck et al. factorised any U into N(N−1)/2 beamsplitter–phase-shifter pairs G[452]; Clements et al. halved the footprint and balanced the loss G[453]; loops reuse one interference point on every time bin G[454]. A six-mode, 15-MZI chip set 100 Haar-random unitaries at fidelity 0.999 ± 0.001 D[455]. Machines on this technology use the network only to sample — single photons after Aaronson and Arkhipov G[456], squeezed light after Hamilton et al. G[268]. Attributes: flying carrier, no gate time or determinism class, loss-dominated error, electro-optic room-temperature control, bulk or photonic-IC fabrication.

Output probabilities are permanents of n×n submatrices of U; exact polynomial-time classical sampling would collapse the polynomial hierarchy, and approximate hardness rests on two conjectures G[456]. With squeezed inputs they are hafnians, also #P-hard G[268]. The mesh entangles modes when fed squeezed light D[270] but has no entangling gate on encoded qubits: linear optics without ancilla photons distinguishes Bell states with at most 50% probability G[457], and universality needs detector feed-forward G[337]. The limits are loss and distinguishability. Transmission compounds with depth: at 0.99 per MZI, a 100-mode Clements mesh passes ~0.37 S[453]. Noisy GBS is efficiently simulable once squeezing, transmission and detector quality satisfy an inequality S[269]; tensor networks are likely efficient when surviving photons scale as √N for N inputs S[458]; at fixed partial distinguishability, simulation cost grows only polynomially with photon number S[459].

Coarse-grained outputs of a linear-optical interferometer are classically reproducible. Any linear statistic of the output — an integer-weighted sum of the photon counts per mode, which covers mode binning, molecular vibronic spectra and the 'most probable bin' one-way function — can be estimated classically to the same additive error as the experiment, in time polynomial in the number of modes, for single-photon and Gaussian inputs P[889]. An advantage or application claim for this mechanism must rest on fine-grained sampling or on a non-linear statistic; hash-function and proof-of-work variants remain open.

Jiuzhang 4.0 answers the MPS attack with efficiency — 92% at the source, 51% end to end — which by the authors' estimate needs bond dimension above 8×10²¹, over 10⁴² years on El Capitan S[175]; validation is on subsystems, with a margin growing with size, not at full scale D[175].

Borealis is fibre and bulk optics: three delays with EOM-driven variable beamsplitters from QUBIG GmbH, 16 TES detectors at 95% behind a 1-to-16 demultiplexer D[270]. Jiuzhang 4.0 cascades three 16-mode interferometers through two delay-loop arrays onto 16 SNSPDs at 93% D[175]. ORCA builds rack-mounted room-temperature systems from telecom-grade fibre components C[464]. Only detectors are cryogenic. No per-element loss of a fielded sampling mesh is published as of 2026-09-26; the 99.8% transmission and >40 dB extinction quoted for Jiuzhang 4.0 belong to its squeezed-light filter D[175].

Borealis sets every loop beamsplitter per 167 ns time bin and samples at 10 kHz between loop stabilisations D[270]. Jiuzhang 4.0 tunes its meshes thermally and loop phases with piezo stretchers; no per-bin electro-optic setting is reported D[175], so the technology's electro-optic attribute fits Borealis best. Readout is terminal; nothing is fed forward. Time multiplexing decouples detectors from modes: 16 detectors read 216 or 8,176 modes D[175], [270]. The heavy I/O is classical: checking a sample exactly means computing hafnians G[268].

The gate slot of the Boson sampler architecture (ph_sampler, goal G1). It requires photon (single photons or squeezed light in the mesh) and ct_eo (phase shifters and switches); it provides no edge — its output is a sample; it replaces nothing, and g_fusion replaces it (fusion vs sampling interferometer): the same optics become a gate only when heralded Bell outcomes (g_fusion) or homodyne results (g_cv) are fed forward. This answers gap G-lointer — samplers had a gate layer with nothing gate-like in it — with an architecture of their own rather than the proposed technology on ph_cv and ph_fusion. Register machines, all primary: Borealis (Xanadu, deployed 2022-06), Jiuzhang 4.0 (USTC, demonstrated 2026-05-13), PT-2 (ORCA, deployed 2024–2026). The clock is a sample period: 36 µs (Borealis), 25.6 µs (Jiuzhang 4.0) D[175], [270].

No gate exists to benchmark; samples are compared with the ideal distribution, and the benchmarks used — low-order correlations, marginal distances — were matched or beaten classically S[461], [463]. The Borealis ✅ and Jiuzhang 4.0 ✅ cells hold against their papers, the latter with the filter correction above. PT-2 stays 🔎: its evidence page gives no architecture C[465]. The one description found is third-party — fibre delay loops with variable beamsplitters, capable of two loops where that paper puts the classically hard regime at three, sampling photon-subtracted squeezed states with a PNR detector P[271] — not the single photons the register records. An ORCA co-authored PT-1 study used one delay line D[466].

ORCA's release of 2025-06-11 names the NQCC system only as "PT Series"; the PT-2 page lists the NQCC testbed as a 2025 deployment. No classical rebuttal of Jiuzhang 4.0 has appeared as of 2026-09-26. Not published: the PT-2 input state in any ORCA primary source.

Money.

• 2024-02-05 · NQCC · £30 M testbed competition, ORCA one of seven winners · per-company value undisclosed · awarded G[467]
• 2024-06-05 · Montana State University · two PT-1 systems, US Air Force-funded · undisclosed · contracted C[468]
• 2025-06-11 · ORCA · NQCC testbed installed under the UK's £121 M initiative · undisclosed · delivered C[469]

Market & supply chain. Of the register's seven machines only ORCA's PT-1 and PT-2 are sold as systems; USTC is academic, and the register marks Borealis superseded by Aurora. ORCA does not publicise most of its fundraising P[470].

IP & standards. No sampler-specific patent count or benchmark standard found as of 2026-09-26.

Roadmaps & track record. NQCC testbeds were due by March 2025 G[467]; ORCA announced its installation on 2025-06-11, a quarter later C[469]. PT-3 is "due for release later in 2026" with three switchable modules, claimed to beat classical solvers on up to 25,000-variable problems R[470].

Strategic reading. A contest between loss and classical simulation, not a route to computation; what lasts are the loops, squeezers and number-resolving detectors that fusion and CV machines also need. If PT-3 reaches the three-loop regime with published loss, ORCA holds the only commercial sampler beyond easy simulation; if not, the technology stays academic.

Confirm if, by 2027-12-31, no classical sampler matches Jiuzhang 4.0 on its subsystem benchmarks and ORCA publishes PT-3 modes, loops and loss with a classical comparison; demote if one matches Jiuzhang 4.0 by then, or PT-3 ships in 2026 without those figures. Open questions. (1) What is the per-element loss of Jiuzhang 4.0's programmable meshes? (2) Does subsystem validation extrapolate to full scale? (3) What does PT-2 inject, and through how many loops? (4) Does any GBS application keep a speed-up at 33–51% transmission? (5) Can a loop sampler become a fusion machine by adding heralding and feed-forward?